Online Appendices to "How productive is workplace health and safety?"

Transcription

1 Online Appendices to "How productive is workplace health and safety?" I. Sebastian Buhai 1, Elena Cottini 2 and Niels Westergård-Nielsen 3 1 Swedish Institute for Social Research, Stockholm University, Sweden; sbuhai@gmail.com 2 Department of Economics, Università Cattolica del Sacro Cuore, Italy;elena.cottini@unicatt.it 3 Department of International Economics and Management, Copenhagen Business School, Denmark; nwn.int@cbs.dk December 2015 Abstract This document contains supplemental descriptive and robustness analyses, mentioned in the main text of our paper Buhai, I.S., Cottini, E. and Westergaard-Nielsen, N. (2015), "How Productive is Workplace Health and Safety", forthcoming at the Scandinavian Journal of Economics. Accessible at

2 Online Appendices to "How productive is workplace health and safety?" A Supplemental material to "How Productive is Workplace Health and Safety?" A.1 Employee vs. employer representative in VOV The VOV 2001 questionnaire is answered both by one safety group representative of the employees ("type-1" respondent) and by one safety group representative of the employer ("type-2" respondent) for each establishment, such that the initial dataset contains two observations for each establishment surveyed. For our reported analysis, we keep only the answer of the employees safety representatives and do not use the second measure, though note that they are fairly highly correlated for the speci c work environment measures (the correlation coe cient is between 0:35 and 0:70 for each of these speci c safety and health measures, with an average across all of them higher than 0:50). Our decision to select type-1 answers is mainly motivated by the fact that their variation is somewhat higher than the ones of type-2, with the latter tending to cluster around "very good" or "good" for most questions. To illustrate the di erence in the variance between the two types with one (extreme) example, consider the answer to the general question concerning the work environment related standard (the correlation between the two measures for this general work environment indicator is only 0.17). Table A.1 presents the answers of both types to the question: "What do you consider the work environment related standard to be at the company?", for observations where both types s answers are nonmissing. We de ne a general work environment ordered variable, taking values that range from 1=very good to 5=very poor. Table A.1: Di erences between types, all plants General work environment Type=1 Type=2 Total N % N % N % very good good not bad poor very poor Total A-1

3 I. Sebastian Buhai et al From Table A.1 it appears clearcut that type 1 answers have more variance than type-2 answers 1, although this discrepancy is less pronounced for any of the speci c work environment indicators. The choice of type-1 versus type-2 answers does not have however any implication on our conclusions: performing our estimations with type-2 answers we obtain results that are qualitatively identical to the ones reported in the paper. A.2 Mono-plant vs. multi-plant rms in VOV Given that we match our datasets via the rm identi er, and that VOV is collected at establishment level, we are compelled to limit our analysis to rms that have a single establishment. How representative is this sub-sample of private Danish sector in terms of geographical and industry distribution, relative to the initial dataset? The two tables below compare their distribution by industries, Table A.2, and respectively the distribution by regions, Table A.3. We note that the mono-plant rms keep largely the same geographical distribution as the plants in the initial sample, and that the only considerable changes are in the case of two industries: real estate, where the proportion of plants decreases from 4.3% of the total sample, initially, to 2.4%, in the working sample, and private rms operating in public administration, defense and compulsory social security category, where the plant percentage decreases from 5% in the initial sample to 0.7% in the working sample of monoplants. Since these are relatively small industries in Denmark, the representativeness of the sample is arguably largely una ected, either industry or location-wise. All our results need to be interpreted, however, with this potential limitation in mind. A.3 Data loss in merging VOV-IDA-REGNSKAB We face some unavoidable sample reduction during the merging procedure, which we describe below: We start with 1962 establishments sampled in VOV 2001 (we have two observations for each of these establishments, corresponding to type-1 and type-2 respondents, as 1 The discrepancy remains the same if we consider only the mono-plant rms, as used in the empirical analysis. A-2

4 Online Appendices to "How productive is workplace health and safety?" Table A.2: Distribution by industries All-plants Mono-plant rms N % N % Agriculture, shing, mining and quarrying Manufacturing Electricity, gas and water supply Construction Wholesale and retail trade Hotels and restaurant Transport, post and communication Financial intermediation Real estate, renting and business activities Public administration, defense and social security Education Health and social work Other community, social and personal service activities Total Table A.3: Distribution by regions All plants Mono-plant rms N % N % Copenhagen Roskilde Vestsjaeland Storstroem Fyn and Bornholms Soenderjylland Ribe Vejle Ringkoebing Aarhus Viborg Nordjylland Total A-3

5 I. Sebastian Buhai et al explained earlier in these Appendices). We need to nd the rm identi er for most of the initial establishments, since these are often included in the dataset only by their name, with that name string often entered only partially, etc. This has been done- via painstaking manual work performed by extremely patient and dedicated research assistants- using an auxiliary business statistics dataset, which matches names to rm identi ers. We were unfortunately not able to retrieve the rm identi er for 490 of the initial establishments. We use only mono-establishment rms in merging VOV to IDA and REGNSKAB since we do not have establishment identi ers in VOV to merge with establishment identi ers in IDA, and since in REGNSKAB we have business account statistics at the business unit= rm level. That leaves us with a sample of 572 rms in the merged VOV-IDA dataset, and 465 rms in the merged VOV-IDA-REGNSKAB dataset. We have less rms in REGNSKAB given the sampling procedure in the construction of that dataset and its reliability only for part of the rms, see also the REGSKAB overview in the data description part of this paper. Remark however, as also stated earlier in the main text of this paper, that using mono-plant rms in productivity analysis is by no means unusual, but rather the established practice when estimating production functions using data originating from both rm (typically, all business account data) and establishment level, see, for instance, Carlsson and Skans (2012). For the production function estimation we use all the available observations in the merged dataset VOV-IDA-REGNSKAB; we end up with smaller sample sizes, given that several of our variables used in the estimation have missing observations. A.4 Descriptive between- rm work environment di erentials In here we detail the descriptive analysis of the health and safety between- rm work environment di erentials mentioned in the data section. Speci cally, we are interested in observable rm and employee aggregate characteristics that correlate with the rms quality of the work environment. Our methodology is similar to, e.g., Osterman (1994), who studied the factors A-4

6 Online Appendices to "How productive is workplace health and safety?" associated with adoption of innovative work practices. Consider the following regression, W E i = + X i + Z i + " i (1) where WE i the indicator of work environment health and safety quality for the ith rm, X i a vector of rm and aggregate-employee characteristics, Z i a vector of work environment practices, and " i an idiosyncratic error term. De nitions and summary statistics for the variables used in our nal speci cation can be found in Table 1 from the main text 2. We estimate simple logit models using both the general and the speci c work environment indicators. In all estimations we report the marginal change in probability of the speci c work environment indicator being 1, given a one unit change in the independent variable. The rst binary outcome model estimated is shown in Table A.4, column (1); our dependent variable is an indicator taking value 1 if the general work environment at the company (as perceived by workers) is "very good" or "good", and respectively 0 if it is "not bad", "poor" or "very poor" 3. The only regressor statistically signi cant at conventional signi cance levels is having undertaken training courses with a work environment content 4, suggesting that rms o ering courses to employees are also more likely to improve their work environment standard, perhaps by rendering their workforce more active and more aware. We stress again that here we aim to emphasize statistical associations, without drawing causal links. Columns (2) to (9) in Table A.4 show estimates for a series of logits, in which the dependent variables refer to speci c work environment problems, with 1 if the speci c condition "is solved" and 0 otherwise. There are few statistically signi cant regressors. 2 Next to the reported speci cations, we have used alternatives with a much wider series of other aggregate employee-characteristics, such as mean and standard deviation, of education, experience, tenure (in the whole rm or per particular employee group, such as managers), as well as proportion of other worker categories, such as "turnover", white-collar, top managers, etc. None of these have statistical power (or a large coe cient magnitude) in explaining the between- rm work condition di erentials, for any of the work environment indicators. 3 We also estimated an ordered probit model with the dependent variable taking 5 values from "very good" to "very poor"; the results are qualitatively identical. 4 Not reported in the table, the age, industry, or geographical location of the rm do not have any explanatory power in this general between- rm work environment di erential. A-5

7 Table A.4: Logit estimates of rm characteristics on work environment indicators, marginal e ects general work heavy lift repetitive work chemical loads noise problems with mental stress internal climate accidents environment young workers (1) (2) (3) (4) (5) (6) (7) (8) (9) proportion female (.119) (.131) (.118) (.038) (.076) (.036) (.121) (.137) (.165) proportion unskilled (.148) (.179) (.165) (.057) (.128) (.065) (.134) (.199) (.139) proportion turnover (.150) (.125) (.146) (.051) (.091) (.032) (.134) (.166) (.141) I. Sebastian Buhai et al proportion managers (.204) (.310) (.302) (.079) (.141) (.101) (.294) (.259) (.235) average education (.019) (.021) (.022) (.012) (.012) (.007) (.023) (.026) (.026) logarithm of rm size (.023) (.026) (.024) (.007) (.015) (.006) (.024) (.026) (.026) courses (.056) (.060) (.059) (.015) (.033) (.016) (.055) (.062) (.052) written policy (.062) (.063) (.067) (.015) (.036) (.015) (.062) (.063) (.058) action plans on work env (.067) (.070) (.079) (.017) (.047) (.017) (.074) (.085) (.064) prioritise work environment (.070) (.080) (.083) (.023) (.053) (.019) (.077) (.096) (.074) Nobs Log-lik Signi cance levels: *** 1%,**5%, *10%; White heteroskedastic-consistent standard errors in parentheses. The estimations also include a constant term, regional and industry indicators, and dummy variables for rm age categories, ie. age 0-5, 5-10, 10-15, 15-20, with the baseline 20+ A-6

8 Online Appendices to "How productive is workplace health and safety?" An unexpected outcome is the importance of the "proportion of managers" 5 in explaining between- rm discrepancies in several of the speci c workplace indicators. In three equations (corresponding to heavy lifting, repetitive work, and mental stress) the coe cient on proportion of managers is positive and statistically signi cant, i.e. a higher proportion of managers in the rm is associated with better work environment in terms of these conditions. The relevance of this indicator extends beyond statistical signi cance, as it also has the largest magnitude relative to other work environment indicators. Having undertaken speci c courses on the work environment correlates positively with the repetitive work and loud noise indicators. The estimates also show that rms with a higher proportion of women are found to have less problems connected with young employees. Finally, prioritizing work environment is found positive and statistically signi cant for solving problems connected to the internal climate 6. The estimated magnitudes of all other covariates found statistically relevant are considerably smaller than that for proportion of managers. The ndings above suggest that there are only a couple of robust rm and aggregateemployee level variables statistically associated with most of the speci c measures of good workplace environment. The most relevant is the proportion of managers, followed by o ering training courses. To less extent, the proportion of females within the rm and prioritizing work environment practice in the rm also help explain across- rm di erences in some of work environment dimensions. Our descriptives could indicate a bene cial e ect of both managerial involvement (proportion of managers) and employee awareness in workplace environment quality improvement. In fact, there exists supportive evidence that these two factors could well be complementary within a rm, e.g., Kato and Morishima (2002), who provide evidence on the association between top-level management and shop- oor employee participation in workplace organization. 5 Other potential proxies for managerial ability, such as means (or standard deviations) of managerial education, age, experience, or tenure, are found not to matter. 6 Not reported, rm age and region are not statistically signi cant for any of the work environment speci c dimensions. However, there are industry di erentials in speci c workplace indicators. For instance, agriculture, is the worst in terms of "heavy load" problems, while chemical loads are worst for the manufacturing category, etc. Since these are not key covariates, we omit them from the tables for the sake of space. A-7

9 I. Sebastian Buhai et al A.5 Analysis using continuous work environment indicators We show that an alternative transformation of the work environment multi-category variables, into continuous, rather than binary indicators, does not change any conclusions. While most literature uses dichotomization of categorical variables in survey contexts like ours, a few studies such as Muñoz de Bustillo et al (2011), Eurofond (2012), or Leschke and Watt (2014) transform them in continuous indicators through mapping onto [0,1]; we repeat our analysis following this strategy, using equal category weights 7. The three tables A.5, A.6 and A.7 below correspond to, respectively, the upper panel of Table 1; the descriptive analysis in online appendix Table A.4; and the main causal analyses in Table 2. Comparing Table A.5 with Table 1 (upper panel), the variance of the continuous indicators is lower than of their binary counterparts. As descriptive analysis, Table A.6 with continuous indicators is very similar to Table A.4., e.g., with proportion of managers strongly associated with several speci c good work conditions. Our estimates from Table 2 appear fully robust to the alternative in Table A.7, with only minor di erences. "Monotonous repetitive work" is now statistically signi cant and quantitatively very important even in the OLS column (1), not only in the FE and GMM, where, just as earlier, "internal climate" also remains statistically signi cant (at lower, but inconsequential, conventional statistical signi cance level) and strongly positive. Table A.5: Descriptive statistics VOV with continuous indicators (corresponds to the upper panel of Table 1 from the main text) variable de nition mean s.d. N VOV continuous indicators general work env. indicator between 0 and 1, increasing in work environment quality heavy lift indicator between 0 and 1, increasing in extent of problem solved repetitive work indicator between 0 and 1, increasing in extent of problem solved chemical loads indicator between 0 and 1, increasing in extent of problem solved noise indicator between 0 and 1, increasing in extent of problem solved problems w young indicator between 0 and 1, increasing in extent of problem solved mental stress indicator between 0 and 1, increasing in extent of problem solved internal climate indicator between 0 and 1, increasing in extent of problem solved accidents indicator between 0 and 1, increasing in extent of problem solved Whereas this indicator uses more information as it does not concatenate categories into one or the other binary dimension, it does make a strong assumption that categories should be equally weighted when mapping onto the continuous [0,1] interval. It is possible to weigh categories unequally, but subject to typically even stronger assumptions (among the three cited studies, only Leshke and Watt, 2014 have also a variant with unequal category weights). Ultimately it is thus not clear whether this method is more e cient than dichotomization, re ected in that most of the literature uses the latter. A-8

10 Online Appendices to "How productive is workplace health and safety?" Table A.6: OLS estimates of rm characteristics on continuous work environment indicators (corresponds to Table A.4 from the Online Appendices) general work heavy lift repetitive work chemical loads noise problems with mental stress internal climate accidents environment young workers (1) (2) (3) (4) (5) (6) (7) (8) (9) proportion female (.070) (.073) (.070) (.047) (.065) (.089) (.108) (.093) (.084) proportion unskilled (.073) (.139) (.152) (.061) (.126) (.197) (.187) (.107) (.117) proportion tunover (.070) (.076) (.080) (.054) (.071) (.094) (.129) (.102) (.080) proportion managers (.099) (.151) (.151) (.101) (.145) (.244) (.235) (.159) (.137) average education (.011) (.011) (.013) (.008) (.014) (.014) (.015) (.015) (.012) logarithm of rm size (.011) (.014) (.014) (.010) (.015) (.020) (.024) (.018) (.017) courses (.028) (.037) (.039) (.026) (.038) (.054) (.056) (.042) (.037) written policy (.028) (.041) (.048) (.027) (.040) (.060) (.062) (.046) (.040) action plans on work. env (.041) (.042) (.057) (.029) (.047) (.076) (.065) (.055) (.055) prioritise work environment (.043) (.052) (.065) (.032) (.058) (.075) (.083) (.058) (.058) Nobs R Signi cance levels: *** 1%,**5%, *10%; White heteroskedastic-consistent standard errors in parentheses. The estimations also include a constant term, regional and industry indicators, and dummy variables for rm age categories, ie. age 0-5, 5-10, 10-15, 15-20, with the baseline 20+ A-9

11 I. Sebastian Buhai et al Table A.7: Augmented production functions with continuous indicators of health and safety conditions (corresponds to Table 2 from the main text) OLS stage FE+OLS 2-stage GMM+OLS (1) (2) (3) 1 st stage K/L.034 (.017).048 (.011).060 (.027) M/L.671 (.026).751 (.022).745 (.061) proportion female (.106) (.053) (.053) proportion unskilled (.111) (.033) (.036) proportion turnover (.130) (.021) (.035) proportion managers (.217) (.075) (.187) average education (.016) (.006) (.008) Nobs 1 st stage Sargan 2 (15)=19.40 (p-value=0.20) LM 1 st order serial corr z=-3.65 (p-value=0.00) LM 2 nd order serial corr z=-0.30 (p-value=0.77) 2 nd stage courses.044 (.035).045 (.035).042 (.034) written policy (.032) (.031) (.030) action plans on work env (.050) (.049) (.048) prioritise work environment (.047) (.049) (.048) heavy lift cont (.083) (.079) (.080) monotonous repetitive work cont..115 (.068).129 (.061).124 (.061) chemical loads cont (.098) (.085) (.085) noise cont (.061) (.052) (.052) problems with young workers cont (.052) (.046) (.045) mental stress cont (.058) (.056) (.056) internal climate cont..022 (.062).076 (.046).082 (.046) accidents cont (.059) (.051) (.050) R Nobs Signi cance levels: *** 1%,**5%, *10%; White heteroskedastic-consistent standard errors in parentheses. Estimations also include a constant term, regional, industry, and rm age category indicators. For the 1st stage FE and GMM regressions in columns (2) and (3), we also control for the interaction between years and industry indicators. The reported GMM here instruments appropriately for K, M, L, and the proportion of managers (there are no qualitative di erences with the case where proportion of managers is not instrumented for). Sargan is a 2 test of overindentifying restrictions; LM is a Lagrange Multiplier test of 1 st and respectively 2 nd order serial correlation in v it, distributed N[1,0] under the null; p-values for the signi cance test of the null hypotheses are reported in parentheses, after the test coe cients A-10